Clinical LOS Radar — Early Length-of-Stay Risk, Done Honestly

An end-to-end clinical-ML case study: from raw MIMIC-IV to a calibrated, leakage-audited, externally-validated model that flags hospital admissions at high risk of a longer stay — early enough for case management to act.

Stack: dbt + DuckDB (SQL transformation & tests) · scikit-learn / LightGBM · MIMIC-IV + eICU (PhysioNet) · Python.

A personal capstone project applying my MIT xPRO Professional Certificate in Data Engineering.

This is an engineering/research portfolio project, not a clinical product. Outputs support operational decisions (bed/case-management prioritization), never treatment.

---

TL;DR — what this project demonstrates

• Framed a real, board-level problem (hospital length-of-stay / throughput, grounded in the AHA Cost of Caring 2025 report) and narrowed it to a single, testable wedge: renal-dysfunction patients at risk of prolonged stay.
• Built a reproducible dbt pipeline (ED-anchored, 0–24h) with hard anti-leakage gates and passing data tests.
• Trained + isotonic-calibrated a model: ROC-AUC 0.736, top-5% precision 82% / lift ~2×, well-calibrated (ECE 0.018).
• Externally validated on eICU (~200 hospitals, retrained recipe): the signal holds at ROC-AUC 0.700 (vs 0.736 on MIMIC) — single-center risk materially reduced.
• Sized the economics honestly (break-even / ROI) and stated what is not proven (the causal LOS reduction — that needs a prospective pilot).

The point of the project is not a model score; it is doing clinical ML in a way you can trust, defend, and ship — leakage discipline, calibration, external validation, and intellectual honesty about limitations.

What's in here. The actual code — the dbt models + tests for the renal-wedge lineage and the Python training / external-validation scripts — plus the write-up, figures, and the metric artifacts that back every number. The data is not included (MIMIC-IV / eICU are credentialed PhysioNet datasets and cannot be redistributed); see docs/data_sources.md.

---

The problem (grounded, not invented)

The AHA Cost of Caring (April 2025) names longer stays, discharge delays, and ED crowding as primary financial pressures; labor is 56% of hospital cost and beds are the constraint. Patients who will occupy a bed for a week are often identified too late. Acute renal dysfunction is one of the report's fastest-growing, costliest utilization drivers — and it is measurable in the first hours from routine labs.

Approach: a narrow, validated wedge

Rather than "predict LOS for everyone" (a crowded, commoditized space owned by Epic, Qventus, LeanTaaS), this project deliberately picks one cohort, one user, one action:

• Cohort: ED→inpatient adults with renal dysfunction in the first 24h (creatinine_max > 1.3 OR BUN_max > 20).
• User: case management / ED operations (operational budget, not FDA-regulated).
• Action: a day-1 prioritized worklist → start post-acute placement & prior-authorization earlier (the documented discharge bottleneck).

A coverage analysis drove a key design decision: creatinine is present in only 20% of admissions at 6h but 66% at 24h (median first creatinine ≈ 14h), so the wedge is anchored at a 24h window — with a "protocolized early creatinine" path proposed for real-time 6h scoring in deployment.

How it's built

raw (MIMIC-IV)
  └─ staging (stg_*)                         dbt views
       └─ encounter_spine_ed_inpatient       1 row / ED→inpatient admission
            ├─ labs_ed_w24h_dynamic          8 labs × {min,max,mean,first,last,delta,range,count,missing}
            ├─ burden_ed_w24h                testing-intensity features
            ├─ context_ed                    age, sex, admission type, ED triage acuity + vitals
            └─ renal_wedge_features          + renal_dysfunction_flag
                 └─ training_renal_wedge     cohort + long_stay label (P75) + subject-level split

Engineering rigor that matters in production:

• Anti-leakage gates as tests: no lab charttime outside [ed_intime, +24h); no discharge/LOS columns in features; no subject in two splits.
• Calibration: isotonic/Platt picked by held-out Brier; reported via ECE (prevalence-aware) instead of a Brier threshold that is mathematically infeasible at this base rate.
• Single source of truth: the trainer reads the dbt mart directly, so the reported numbers and the pipeline can never silently drift apart.
• Prevalence-aware evaluation: operating points (top 5/10/20%) report precision + lift + the recall ceiling, not an impossible recall target.

Results

Population	ROC-AUC	Top-5% precision / lift	Top-10% precision / lift	ECE	base
MIMIC-IV (ED→inpatient, single center)	0.736	82% / 1.95×	76% / 1.80×	0.018	41%
eICU (ICU, ~200 hospitals) — external	0.700	60% / 2.37×	55% / 2.17×	0.008	25%

Retrained on the multi-center eICU population, the renal-cohort signal holds (0.736 MIMIC → 0.700 eICU), materially de-risking the single-center concern. Within the cohort the prolonged-LOS rate is 41% vs 25% baseline (1.64× enrichment), so the wedge concentrates risk before any model is applied. Every figure above is reproduced from the committed artifacts in this repo (docs/renal_wedge_model_metrics.json, docs/eicu_external_validation.json).

Economics (illustrative, honest): at one large center the renal cohort sits on ~$30M/yr of excess bed-days; the early-creatinine protocol costs ≈ 0.23% of that pool, so a 5–10% LOS reduction would repay it many times over. See docs/wedge_onepager_renal.md.

What this does NOT prove (read this)

• That earlier flagging causally reduces LOS. We proved the cohort is identifiable, rankable, and generalizes; whether earlier case-management action shortens the stay is unproven and is exactly what docs/pilot_design_renal_wedge.md is designed to measure. Prior "generic" LOS interventions in the literature often failed — likely because they didn't target the discharge bottleneck.
• Anything clinical. The flag prioritizes a scarce operational resource; it does not recommend treatment.

Reproduce

Data is not included (obtain MIMIC-IV / eICU via credentialed PhysioNet, build a DuckDB at data/processed/mimic.duckdb). Then:

pip install -r requirements.txt

# 1) build the dbt wedge lineage + run the anti-leakage / integrity gates
cd cdsp
cp profiles.yml.example ~/.dbt/profiles.yml      # then point `path` at your DuckDB
dbt build --select +training_renal_wedge
cd ..

# 2) train + isotonic-calibrate + write metric artifacts
python scripts/train_renal_wedge_model.py --from-mart --write

# 3) external validation on eICU (writes docs/eicu_external_validation.*)
python scripts/validate_eicu_renal_wedge.py --eicu-dir data/raw/eicu

Repo map

Path	What
cdsp/models/	dbt models — staging → spine → features (labs_ed_w24h_dynamic, burden_ed_w24h, context_ed) → renal_wedge_features → training_renal_wedge
cdsp/tests/	anti-leakage + integrity gates (gate_labs_ed_w24h_anti_leakage, gate_training_renal_wedge_subject_split_leakage, …)
scripts/train_renal_wedge_model.py	train + isotonic calibration + prevalence-aware operating points
scripts/validate_eicu_renal_wedge.py	external validation on eICU
scripts/wedge_feasibility_renal.py, diagnose_creatinine_coverage.py, wedge_roi_breakeven.py	cohort sizing, coverage analysis, ROI / break-even
docs/wedge_onepager_renal.md	business one-pager (problem → wedge → economics)
docs/radar_output_contract.md	output contract + prevalence-aware targets
docs/pilot_design_renal_wedge.md	prospective pilot design + theory of change
docs/anti_leakage_phase3.md	how leakage is prevented
docs/*_metrics.{json,md}, docs/eicu_external_validation.{json,md}	the artifacts behind every number above
assets/	figures

---

Scope & data governance

• No patient-level data is included — code, documentation, figures, and aggregate metric artifacts only.
• MIMIC-IV / eICU accessed under credentialed PhysioNet terms; used for research only.
• Operational decision support, not medical advice; not an FDA medical device as scoped.